What’s in a Picture? Evidence of Discrimination from Prosper.com*

Devin G. Pope The Wharton School

University of Pennsylvania

Justin R. Sydnor Department of Economics

Weatherhead School of Management Case Western Reserve University

This Draft: September, 2008

Abstract

We analyze discrimination in a new type of credit market known as peer-to-peer lending. Specifically,

we examine how lenders in this online market respond to signals of characteristics such as race, age, and

gender that are conveyed via pictures and text. We find evidence of significant racial disparities; loan listings

with blacks in the attached picture are 25 to 35 percent less likely to receive funding than those of whites

with similar credit profiles. Conditional on receiving a loan, the interest rate paid by blacks is 60 to 80 basis

points higher than that paid by comparable whites. Though less significant than the effects for race, we find

that the market also discriminates somewhat against the elderly and the overweight, but in favor of women

and those that signal military involvement. Despite the higher average interest rates charged to blacks,

lenders making such loans earn a lower net return compared to loans made to whites with similar credit

profiles because blacks have higher relative default rates. This pattern of net returns is inconsistent with

theories of accurate statistical discrimination (equal net returns) or costly taste-based preferences against

loaning money to black borrowers (higher net returns for blacks). It is instead consistent with partial taste-

based preferences by lenders in favor of blacks over whites or with systematic underestimation by lenders of

relative default rates between blacks and whites.

Contact Pope at dpope@econ.berkeley.edu and Sydnor at justin.sydnor@case.edu.

*We thank David Card, David Clingingsmith, Stefano DellaVigna, Jonathan Guryan, Erzo Luttmer, Nicola Persico, Jim Rebitzer,

Stephen L Ross, Heather Royer, Jonathan Skinner, Nick Souleles, Betsey Stevenson, Justin Wolfers and seminar participants at the NBER Summer Institute, Case Western Reserve University, and the University of Pennsylvania for helpful comments and suggestions. All errors are our own.

1

There is a long history within economics of studies attempting to understand discrimination in a variety

of markets. Much of this interest stems from concerns that because of discrimination, certain groups – for

example, blacks and women – may not enjoy the same access to markets and opportunities as their

counterparts. Theories of discrimination usually fall into one of two classes: statistical discrimination

(Phelps, 1972; Arrow, 1973) or taste-based discrimination (Becker, 1957).1 Accurate statistical

discrimination is economically efficient for the decision maker, while taste-based discrimination stems from

an animus toward one group and is often costly to the decision-maker. Because costly discrimination may be

driven out of competitive markets, and because these different theories often lead to different policy

recommendations, understanding the extent to which observed discrimination is consistent with these

theories is an important goal. However, it is often difficult to test for discrimination in markets2 and

generally even harder to assess the different theories of discrimination.3

This paper examines discrimination in a new type of credit market known as peer-to-peer lending.

Specifically, we study data from the website Prosper.com, a leader in online peer-to-peer lending in the

United States. Peer-to-peer lending is an alternative credit market that allows individual borrowers and

lenders to engage in credit transactions without traditional banking intermediaries. While still small, these

markets are growing quickly and may represent an important niche, especially in the area of consumer-debt

consolidation.4 Websites like Prosper aggregate small amounts of money provided by a number of

individual lenders to create moderately-sized, uncollateralized loans to individual borrowers. In order to

request funding, borrowers in these markets create a loan listing that resembles auction listings for goods on

websites like eBay. Like most standard credit applications, this listing displays desired loan parameters and

reports information from the prospective borrower’s credit profile. Unlike typical credit applications,

however, borrowers may include optional and unverified personal information in their listings in the form

of pictures and text descriptions. These pictures and descriptions often provide potential lenders with

1 For other literature on theories of discrimination see Aigner & Cain (1977) and Lundberg & Starz (1983).

2 See Altonji and Blank (1999) and Blank, Dabady, and Citro (2004) for reviews of empirical work on assessing discrimination in labor markets, and Ross and Yinger (2002) for a similar review in credit markets with a focus on mortgage lending. 3 A few notable papers that have used clever empirical methodologies to examine statistical discrimination vs. taste-based discrimination include Altonji and Pierret (2001), Knowles, Persico, and Todd (2001), Levitt (2004), Antonovics and Knight (2004), and Charles and Guryan (2007). 4 See Freedman and Jin (2008) for an analysis of the evolution of the Prosper market and the profitability of loans on Prosper.

2

signals about characteristics such as race, age, and gender, that anti-discrimination laws typically prevent

traditional lending institutions from using.

Our first research question focuses on the determinants of access to credit in the Prosper marketplace,

and in particular on how signals from pictures about characteristics, such as race, age, and gender, affect the

likelihood of receiving loan funding and the interest rates borrowers pay. In the language of the legal

literature we test for “disparate treatment” of certain groups by estimating whether they are treated

differently than their counterparts who are similar on other dimensions.5 Our empirical approach uses

observational market data.6 The typical problem with this type of analysis is the potential for omitted

variable bias.7 Fortunately, however, our data set includes all of the information that lenders see when

making their decisions. Prosper.com generously provides a data set that contains all of the information from

loan listings created on the site, including links to pictures included with the listings. In order to conduct

the analysis, we systematically coded variables from pictures and text descriptions for over 110,000 loan

listings that were created on Prosper.com between June 2006 and May 2007.

The empirical analysis reveals significant racial discrimination in this market. Compared to the response

to otherwise similar whites, we estimate that listings with blacks in the picture are 2.4 to 3.2 percentage

points less likely to be funded. Compared to the average probability of funding, 9.3%, this represents an

approximately 30% reduction in the likelihood of receiving funding. A range of estimation techniques –

OLS regressions, Logit estimation, and propensity-score analysis – and numerous robustness checks and

5 The other important definition within the legal literature is “disparate impact,” which arises when decision-makers do not explicitly account for characteristics such as race and gender, but use variables that are highly correlated with these characteristics. See Ross and Yinger (2002) for a discussion of disparate treatment vs. disparate impact with a focus on discrimination in credit markets. 6 This observational-market-data approach is similar to that used in the influential studies of redlining and racial discrimination in mortgage lending by the Boston Federal Reserve (Munnell, Tootell, Browne, and McEneaney (1996) and Tootell (1996)). 7 Audit studies and field experiments are an important alternative technique for examining the existence of discrimination (specifically disparate treatment) in a range of markets. For instance, in a very influential paper Bertrand and Mullainathan (2004) study racial discrimination in the labor market by randomly assigning race to fictitious resumes and find that resumes with black-sounding names are less likely to receive a call-back for an interview. Examples of audit studies include Turner et al. (2002) on mortgage lending, Turner et al. (1991) on the labor market, and Ayers and Siegelman (1995) on automobile purchases. By manipulating the race or gender of applicants for jobs or loans, these types of studies are able to identify clean causal links between group status and treatment without concerns of omitted variables or the correct empirical specification. On the other hand, because they usually lack any ex-post performance data, with audit or field-experiment approaches, it is generally hard to assess different theories of the sources of discrimination. Heckman (1998) argues that the audit approach over-states the importance of discrimination, arguing that while some employers, salesmen, or lenders may discriminate, that minorities will seek out those who do not, thereby lessening the impact of the discrimination. It is worth noting that the discrimination we find on Prosper.com is at the market level.

3

alternative cuts of the data reveal very stable effects of race on the likelihood of funding. This

discrimination against blacks in the lending decision is also reflected in the interest rates these borrowers pay

conditional on receiving a loan; their interest rates are 60 to 80 basis points higher than those of whites with

similar credit profiles.

While smaller and less robust than the results for race, we find a number of other interesting market

responses to the information in pictures and text. For instance, the market discriminates somewhat against

the elderly and significantly overweight, but in favor of women and those that signal military involvement.

The market also favors listings where the borrower expresses a desire to pay down credit-card debt (the

most popular stated loan purpose) over credit requests for other purposes, such as loans for business

expansions or automotive repairs/purchases.

It is perhaps somewhat surprising that we find evidence of discrimination in this market. Because the

pictures and descriptions are optional and unverifiable, a natural prediction would be that the market would

respond little to this type of “cheap talk”. Yet the fact that borrowers include a wide variety of pictures and

the market responds to those signals, suggests that the information is not treated as cheap talk in the market.

In fact, we find the Prosper market responds negatively to listings that do not include a picture. Another

reason that the finding of racial discrimination might be somewhat surprising is that lenders are given a wide

range of information about each borrower’s credit profile, including credit grade, debt-to-income ratio, and

a measure of income. However, we find that lenders respond to signals about race above and beyond this

wealth of credit information.

Given that we find discrimination in this market, an obvious question is whether this discrimination is

efficient for lenders – i.e., are these differences consistent with lenders engaging in accurate and

economically efficient statistical discrimination? Because of the availability of data and the nature of the

market, we can address this question using loan-performance data.8 A unique feature of the Prosper market

is that it operates as an auction that allows interest rates to be bid down below an initial rate set by the

8 Exploring theories of discrimination – i.e., statistical discrimination vs. taste-based discrimination – is generally quite difficult. In many settings there is no ex-post performance data available. Even when performance data is available, it may not be informative because decision-makers use a threshold cutoff for decisions such as loan approval. For example, see the critiques of the use of default analysis to assess theories of discrimination in mortgage lending that appeared in the May, 1996 edition of Cityscape, especially articles by John Yinger, George Galster, Stephen Ross, and John Quigley.

4

borrower, if enough lenders find a loan attractive. The basic intuition behind the analysis, then, is that if

lenders care only about the net return of a loan (adjusted for expected default), funds will flow to loans that

are attractive given the observable information to lenders. This process should adjust their interest rates and

equalize expected returns. If the market correctly incorporates characteristics from pictures and text when

assessing creditworthiness, accurate statistical discrimination will result in funded loans that have equal

average net returns irrespective of the listing characteristics. On the other hand, if taste-based

discrimination is the sole cause of disparate treatment in the market, loans made to the group subject to

negative discrimination should have higher net returns ex post.

The comparison of the net return on loans made to blacks and otherwise similar whites is striking. The

estimated average net return on a dollar from investing in a loan from a black borrower is 7.3 to 8.6

percentage points lower over a three-year period. Although blacks are discriminated against in the lending

process, the higher interest rates that they pay are not enough to account for their greater propensity to

default. This runs counter to the predictions of both accurate statistical discrimination (i.e., equal net returns)

and taste-based animus against blacks (i.e., higher net returns on loans to blacks).9

How can we reconcile the evidence of discrimination against blacks in the lending process with the fact

that their loans result in lower net returns? The evidence is consistent with a combination of accurate

statistical discrimination against blacks coupled with taste-based discrimination against whites. But such an

interpretation runs counter to intuition and to previous literature, which rarely concludes that there is a

taste-based preference against whites. We discuss the interpretation of these results in detail at the end of

the paper. Perhaps the most likely interpretation is that lenders understand the correlations between race

and important characteristics for predicting default that they cannot perfectly observe, such as education

9 After we had gathered our data and were conducting our analysis, we learned of a working paper by Ravina (2008) that conducts

an analysis similar to ours but uses a smaller sample of loans (one month of loan listings on Prosper relative to the twelve months used in our analysis). Ravina’s strongest findings are for the effect of beauty, and she finds that more beautiful people are more likely to receive funding. In contrast we find little effect of our attractiveness measure, which we attribute to her more precise coding of beauty. Our race coding, however, is quite accurate and we find a number of differences in our results for race. While Ravina finds that blacks pay higher interest rates conditional on funding (consistent with our results), her estimates do not show a difference in the probability of funding related to race. Furthermore, Ravina concludes that there is no evidence of differential default rates between loans made to blacks and whites. Most of these differences can likely be attributed to the large differences in sample sizes, as the standard errors on her estimates are large and cannot reject our point estimates for any of the estimations even though our results are highly statistically significant.

5

and social-support networks, but they fail to fully appreciate the strength of these correlations or the

importance of these unobservable factors in predicting default.

The remainder of the paper proceeds as follows: Section I describes peer-to-peer lending and the

dynamics of the Prosper marketplace. Section II develops a simple model of the Prosper market that

motivates and focuses the analysis of discrimination in this market. Section III describes the data made

available by Prosper.com and our process for coding information from pictures and text. Section IV

presents our empirical results, focusing first on the probability of obtaining a successful loan and then

turning to estimates of the net return (to lenders) of loans made to different groups. We conclude the paper

in Section V with a discussion of the interpretation of our results and their relationship to and implications

for the literature on theories of discrimination.

I. Institutional Background of Online Peer-to-Peer Lending

Online peer-to-peer lending encompasses a range of new and expanding markets that allow individual

borrowers and lenders to engage in credit transactions without traditional intermediaries such as banks.

These markets are small but growing quickly: the U.S. peer-to-peer market grew from an estimated $269

million in outstanding loans in 2006 to $647 million in 2007.10

Part of the appeal of peer-to-peer lending is that it offers lower overhead and the ability to cut out the

bank or “middle man”. Of course, there are many reasons why banks and other credit agencies have

historically been the primary source for personal loans. Prosper has addressed some of the most important

advantages of traditional lending institutions, including enabling individuals to diversify their peer-to-peer

lending portfolio and providing individuals the sort of credit-profile information that until recently has been

the purview of banks and other large lenders. Naturally, it is questionable whether individuals have the

sophistication and training to make efficient use of this credit information in the way banks can. On the

other hand, peer-to-peer markets provide lenders with a wealth of personal and contextual information

about borrowers that traditional intermediaries do not use and are often explicitly barred from using by anti-

10 This information comes from an article entitled “How to Use Peer-to-Peer Lending Sites to Borrow Money,” that appeared on foxbusiness.com on Monday, January 28 2008, and cites its source as the research firm Celent. According to the article, Celent projects the market to grow to a total of $5.8 billion by 2010.

6

discrimination laws. This extra information may be a source of advantage for peer-to-peer markets.

Ultimately, because they are so new, it is still too early to know whether peer-to-peer credit markets will

actually succeed, but they are an intriguing alternative to traditional credit markets and are attracting both

borrowers and lenders.

Details of Prosper.com. Our analysis focuses on the Prosper.com marketplace. Started in

February, 2006, Prosper is somewhat similar to auction sites such as eBay, except that instead of bidding on

or listing a consumer item, individuals bid on or list personal loans. All loans in this market are

uncollateralized and have three-year terms with a fixed repayment schedule. Individuals wishing to borrow

money create a listing that lasts for a pre-specified length of time, usually between 7 and 14 days. The listing

includes the amount of money requested (up to $25,000), the maximum interest rate the borrower is willing

to pay, credit information obtained by Prosper via a credit check, and voluntarily provided (and unverified)

information, such as pictures and descriptions of what they plan to do with the money. Lenders browse the

various listings and bid on specific loans by committing a portion of the principal (minimum of $50) and

setting the lowest interest rate at which they are willing to provide those funds. The loan gets funded if and

only if the total amount of money bid by lenders covers the size of the requested loan. Lenders get priority

for the loan based on the minimum interest rate they are willing to accept, with low-rate bids getting higher

priority. If enough lenders bid on the loan, the final interest rate on the loan can be bid down from the

maximum interest rate initially set by the borrower; the final rate is determined by the lowest reservation rate

set by a bidder who does not get to fund a portion of the loan.11 Prosper makes money by charging closing

costs of 1-2% of the loan amount to borrowers and 0.5-1% to lenders.

An example may help clarify the market dynamics. Imagine that a borrower requests a $5,000 loan

and is willing to pay a maximum annual interest rate of 10%. For simplicity, assume that all potential

lenders will bid the minimum funding size of $50. It then takes 100 lenders to fund the $5,000 loan. Each

of these lenders enters a reservation interest rate when they bid, which is the lowest interest rate they are

11 Although we (and Prosper) use the term “lenders” to refer to the individuals making bids for the loan, technically speaking the loan contract is between the borrower and Prosper. So borrowers do not have to make separate repayments to each lender, but rather simply repay Prosper based on the final interest rate for their loan. Prosper allocates the repayments to the individual lenders based on the portion of the loan funds they provided.

7

willing to accept. If there are exactly 100 lender bids, the $5,000 loan will fund at an interest rate of 10%.

However, if more than 100 lenders bid on the loan, the final interest rate would be determined by the 101st

lowest reservation interest rate. The 100 bidders with the lowest reservation interest rates would each

provide $50 for the loan and would be entitled to 1/100th of the repayments made by the borrower over the

three-year term.

There is substantial information available to individuals who are interested in bidding on loans.

Lenders see the parameters of the loan: its size, the ending time of the listing, the total amount that has been

funded through bids by other lenders, the history of bids on the listing, and the current interest rate, which

is either the maximum rate the borrower will accept or (for fully funded loans) the rate to which the loan has

been bid down. Other than these loan parameters, perhaps the most important information available to

lenders is a credit profile for each borrower obtained by Prosper through a standard credit check. Prosper

obtains an Experian credit score and provides lenders with a credit grade (e.g., AA or B) for each borrower

using bins of credit score.12 The cutoffs for the different credit grades are found easily on the Prosper

website, but lenders do not see borrowers’ exact credit scores. Lenders also see a host of other information

commonly found on credit reports, including delinquencies, revolving credit balance, and bank-card

utilization.13 Potential borrowers also supply information about their employment status, occupation

(chosen from a list), and income. The income borrowers report is also used by Prosper to create a debt-to-

income ratio that is prominently displayed on the listing pages. This debt-to-income ratio is calculated by

dividing the borrower’s self-reported income by his or her debt burden (excluding housing) as reported by

the credit check, and includes the value of the Prosper loan the borrower is requesting. Prosper does not

verify the employment, occupation, and income information when loan listings are created, but does verify

this information for some borrowers once the loan becomes fully funded and before the money is

12 Credit grade bins include the following: AA (760 and up), A (720-759), B (680-719), C (640-679), D (600-639), E (560-599), and HR (520-559). Individuals with a credit score below 520 are not allowed to create a loan listing. 13 Additional information in the credit profile includes, the numbers of public records in the last year and last ten years, the number of inquiries in the last six months, the date of the borrower’s first credit line, the numbers of current, open, and total credit lines.

8

disbursed. The final piece of financial information provided to lenders is an indicator for whether the

borrower is a homeowner or not.14

In addition to this financial information borrowers can include supplemental material in their listing

consisting of: a) a picture with their listing, b) a one-line description for the loan, and c) a separate longer

description, where borrowers are encouraged (by Prosper) to describe what they plan to do with the money

and why lenders should consider their request. None of the information in these pictures or descriptions is

verified by Prosper or verifiable by lenders.

Prosper also incorporates additional social components through the use of borrower (and lender)

groups. Borrower groups are generally organized around some sort of theme (e.g., alumni of a particular

university) and include a rating. The group rating is affected by the repayment activities of its members so

that group membership provides extra social pressure to repay loans.

Other than social pressure and conscience, the primary incentive for a borrower to repay the

uncollateralized loan is the impact that default can have on the borrower’s credit. If a borrower fails to

repay the loan, Prosper reports the default to the credit-scoring agencies and turns the loan over to a

collection agency that attempts to recover some money.15 Ultimately the penalties to a borrower from

defaulting on a loan in this market are similar to those of failing to repay a credit card.

II. Model

This section develops a simple model of the peer-to-peer lending market based on the nature of the

loan-auction process in the Prosper.com marketplace. We are interested in understanding how potential

lenders respond to the information that borrowers reveal through their pictures and descriptions.

The lending market consists of prospective borrowers indexed by i and potential lenders indexed by j.

Assume that there are two groups of potential borrowers in the market whose only observable difference is

membership in either a majority or minority group (e.g., white and black). Also, for simplicity, assume that

14

More recently, Prosper.com has initiated several new features that may allow lenders to receive additional information. For example, recently Prosper.com added a feature that allows lenders to converse with potential borrowers via email. However, this and several other features post-date the data that we work with in our paper. 15 Any money recovered by the collection agency is repaid to the individual lenders in proportion to the amount of the loan they funded.

9

Eq.(1a)

all borrowers request loans of the same size L, and that lenders diversify their portfolio by offering to fund

only 1/nth of any loan they find attractive. In other words, it takes n lenders to fund a loan. Individual

borrowers set a maximum interest rate they are willing to pay for a loan, denoted by 𝑟 𝑖 . Lenders “bid” on

loans by stating the minimum interest rate they require to be willing to lend to the borrower. A prospective

loan funds if the borrower’s maximum interest rate is at least as large as the bid from the lender with the nth

lowest reservation rate. The market is structured such that the final interest rate on a loan that funds is

equal to the (n+1)st bid (again ranked low to high); so lenders find it optimal to bid their true reservation

interest rate for any loan.

The lenders’ reservation interest rates are determined by perceived probabilities of default and may

additionally be influenced by taste-based preferences for one group over another. Let the probability of

loan default for borrowers from the majority group be p and the probability for minority borrowers be (p

+), where for simplicity we assume that (p +) < 1. The parameter represents any true (statistical)

difference in average default rates between the groups. The expected net return on a loan with interest rate r

is then:

1 − 𝑝 1 + 𝑟

for majority loans and

1 − 𝑝 − 𝛾 1 + 𝑟

for minority loans.

Lenders without taste-based preferences for a particular group set reservation interest rates for loans

from each group such that the perceived expected return is equal to the rate of return from the outside

option. With the rate of return from the outside option set at zero (for simplicity and without loss of

generality), this results in the following reservation interest rates:

𝑟 =𝑝

1 − 𝑝,

for loans to the majority, and

Eq.(2a)

Eq.(1b)

10

𝑟 𝑚 =𝑝 + 𝛾

1 − 𝑝 − 𝛾= 𝑟 +

𝛾

1 − 𝑝 − 𝛾 (1 − 𝑝) ,

for loans to the minority. If there is taste-based animus toward the minority group, lenders require

additional compensation for lending to the minority, which adjusts the reservation interest rate for loans to

minorities so that:

𝑟 𝑚 = 𝑟 +𝛾

1 − 𝑝 − 𝛾 (1 − 𝑝)+ 𝛿 ,

where quantifies the degree of taste-based animus against ( > 0) or in favor of ( < 0) the minority

group.

Allowing for randomness in the outside option for individual lenders, we have individual reservation

interest rates: 𝑟 𝑗 = 𝑟 + 휀𝑗 and 𝑟 𝑗𝑚 = 𝑟 𝑚 + 휀𝑗 , where the j are randomly drawn from a distribution with

non-negative support. We can order the lenders from smallest to largest reservation interest rates, such that

j = 1 denotes the lender with the lowest reservation rates. A loan will fund if there are at least n lenders

willing to fund the loan, which implies that 𝑟 𝑖 𝑟 𝑛 for a majority borrower and 𝑟 𝑖 𝑟 𝑛𝑚 for a minority

borrower. If the loan funds, the final interest rate on the loan will be 𝑟 𝑛+1 or 𝑟 𝑛+1𝑚 for majority and minority

loans respectively. The difference between the reservation rates that a lender sets for minority loans and

majority loans is:

𝑟 𝑗𝑚 − 𝑟 𝑗 =

𝛾

1 − 𝑝 − 𝛾 (1 − 𝑝)+ 𝛿.

This difference is increasing in both (the true difference in default rates between minority and majority

borrowers) and (the degree of taste-based animus against minority borrowers). Since the final interest rate

Eq.(2b)

Eq.(2c)

Eq.(3)

11

on a funded loan is determined by the preferences of the lender with the (n+1)st bid, contingent on actually

getting funded, the difference in interest rates between minority and majority borrowers is also determined

by Equation (3). Thus, either taste-based discrimination (non-zero ) or accurate statistical discrimination in

the presence of average group differences (non-zero ) will lead to differences in both the likelihood of

funding and the interest rate conditional on funding between minority and majority borrowers who set the

same maximum interest rates.

If the interest-rate cutoffs for the two groups are determined solely by accurate statistical discrimination

(Equations 2a and 2b), then the difference in the final interest rates between funded minority and majority

loans equalizes the expected net returns on loans to the two groups. On the other hand, if lenders have

taste-based preferences for one group (i.e., 0), the difference in the expected net return on loans made to

the two groups will be equal to .

One question that often arises in the literature on discrimination is whether observed discrimination

should be interpreted as an average or a marginal result. The empirical observations in this paper give the

average level of discrimination at the market equilibrium. However, since each listing goes through an

auction process to determine funding and interest rates, each listing is essentially on the margin. This

market dynamic allows us to assess theories of discrimination using loan-performance data in a way that

studies of other credit markets typically cannot. The basic problem in other settings, for instance the

mortgage market, is that lenders may set a creditworthiness cutoff for access to loans at a fixed interest rate.

In such a setting, if the econometrician does not observe the underlying creditworthiness of the applicant

and different groups have different distributions of creditworthiness above the cutoff, analysis of average

returns may suggest that lenders engage in taste-based discrimination even when they use the same marginal

cutoff for each group (Ross 1996, 1997). The individualized loan-decision process and wealth of data in the

Prosper market overcome these common problems.

Most discussions of theories of discrimination assume, as we have here, that decision-makers have

beliefs about the majority and minority that are on average accurate. Another possibility, however, is that

lenders have biased beliefs in which they systematically misperceive the relative probability of default

12

between the groups. The literature on discrimination treats this possibility in a confusing manner:

sometimes ignoring it, sometimes grouping it under the heading of “statistical discrimination”, and

sometimes grouping it with taste-based preferences under a heading of “prejudice”. In order not to confuse

our discussion, we treat biased beliefs as a third case, distinct from accurate statistical discrimination and

taste-based discrimination. However, the effects of biased beliefs and taste-based preferences are

observationally equivalent and could both be included in the parameter under the heading of prejudice.

A final word on the model is in order. The model takes the maximum interest rate that a borrower sets

as exogenous. Although the bid-down process should in theory allow borrowers to set their maximum

interest rate at their reservation rate, if there is uncertainty about that process, in practice different groups

might set different interest rates because of group membership. For instance, suppose that black borrowers

anticipate racial discrimination and thus set higher interest rates in order to increase their chances of

funding. Since our empirical specifications for the probability of funding include controls for the rate that

the borrower sets, as long as different groups do not set completely divergent interest rates, we can isolate

the ceteris paribus effect of group membership on the likelihood of funding. If, however, one were interested

in a different question about the average experience of discrimination in the market, this type of anticipation

of discrimination might make it appear as though a group is more likely to receive funding and is

discriminated against less than they actually are.

III. Data

Data Overview. Prosper.com generously makes its data available to academics and prospective

lenders. Data are available for every loan listing since the inception of the website. The data include all of

the information seen by lenders when they make their lending decisions, as well as the outcome of the listing

(i.e., funded or not). Demographic and other information about lenders is not available.

Figure 1 graphs the number of requested loan listings made on the website over time. The number of

listings grew quickly after Prosper’s official launch in February, 2006, reaching 5,000 requested loans per

month by May, 2006 and rising to over 10,000 listings per month by January, 2007. The number of loans

13

that actually get funded, however, has risen much more slowly. Of the 203,917 loans requested between

February, 2006 and November, 2007, 16,395 were funded (8.04%), with lenders providing a total of

$101,913,173 in funds (mean $6,216 per loan) to borrowers. The large number of loan requests that go

unfunded motivates our interest in understanding how the market chooses which loans to fund.

The vertical bars in Figure 1 highlight the time-period we study in this paper. We focus on all loans

that were listed during a one-year period in the Prosper market from June 2006 through May 2007, which

leaves out the first few months of the market and ensures that we have at least seven months of repayment

data for any loan made. Table 1 provides a series of summary statistics for the loan listings that occurred

during the sample year. The columns in the table provide information about the full sample of loan listings

and the subset of listings that actually funded. During this year, there were 110,333 distinct loan listings, of

which 10,207 (9.3%) funded. The average requested loan size for all listings was $7,154 and was $5,930 for

the funded listings, revealing that during this period just over $60 million in funds were lent through

Prosper. On average borrowers set a maximum interest rate of 17% on loan listings. Among the loans that

actually funded, however, borrowers set a maximum interest rate of 20% and had an average final interest

rate (after bid down) of 18%. It is also worth noting that 43% of loans are specified as loans that “fund

immediately”. Rather than letting lenders bid down the interest rate, borrowers of these loans request that

the loan is processed as soon as funding is available at the initial interest rate that was specified.

Credit Data. Prosper uses eight credit grades in their credit-scoring process. The majority (54%) of

the requested loans are made by individuals who fall into Prosper’s “high risk” (HR) credit grade with credit

scores from 520-559. Listings with these credit grades are less likely to fund, however, and represent only

20% of the funded listings. Listings from individuals with the best credit grades (AA and A), who have

credit scores above 720, each make up 3% of the total listings, but are more likely to fund and make up 10%

and 9% of the funded listings, respectively. The average debt-to-income (DTI) ratio of 63% for those

requesting loans also confirms the poor credit situation of the typical prospective borrower. Those who

actually get loans are in a better financial situation, but still have rather high average DTI at 39%.

Coded Data from Pictures and Text. To obtain data from pictures and descriptions, we employed

a number of undergraduate research assistants to systematically code up the information in the borrower’s

14

picture (if included) and the borrower’s one-line description of the loan for all 110,333 loan listings on

Prosper during the sample year. These assistants were paid a simple piece-rate per listing, and were informed

that we would randomly check approximately 10% of their entries for accuracy. On the rare occasion that

one of the coders made a large number of errors, he or she was asked to redo the coding and was not paid

until a thorough accuracy check was performed. The coders were not told about the underlying hypotheses

of the research, and importantly did not see any of the parameters of the loan listing other than the picture

and one-line description while coding.16

The coders used the text descriptions to classify the purpose of the loan. This categorization provides

an interesting picture of why borrowers are asking for money on Prosper.com. The categories for these

purposes are listed in Table 1 and were chosen as the most frequent and important categories after a review

of 750 loan listings. Around 30% of the listings used a description that stated the purpose of the loan as

being some form of debt consolidation (e.g., “consolidating credit card debt”, “pay down debt”, and

“paying off credit cards”). This is consistent with media reports that often stress the potential value of the

peer-to-peer credit market as a way out of credit-card debt. Another popular category (10% of all listings) is

business or entrepreneurship loans (e.g., “expanding my successful small business”, “a new truck for

landscaping business”). Smaller percentages communicated that they needed money for education expenses

(3%), medical/funeral expenses (3%), home repairs (2%), automobile purchases (2%), automobile repairs

(1%), or to pay back taxes (1%). A sizeable number of listings (34%) did not fall into these main categories

(e.g., “need help”) and were coded under a category of unclear/other.17 Interestingly, and in contrast to the

financial information, the distribution of loan purpose is quite similar between the funded listings and the

full sample of listings, suggesting that the stated purpose of the loan is not a particularly important

determinant of loan funding.

We hand coded only the text in the one-line description and not in the longer description that

borrower’s provide with their loan. The costs to hand coding information from these longer descriptions

were simply prohibitive. Instead we ran the longer text descriptions through a simple text-analysis program

16 Copies of the coding protocols that we gave to the research assistants are available on request. 17 Approximately 6% of listings included multiple reasons for wanting the loan within their descriptions (e.g., “pay off a car loan and attend a family reunion”), and we coded these multiple-purpose listings under a separate category.

15

that outputs the number of characters, words, and sentences in the text, an index of readability based on the

average word-length and average sentence-length, and the percent of words that are misspelled.18 These

text-analysis variables are slightly correlated with measures of creditworthiness and picture characteristics19,

and we include controls for them throughout our analysis.

Turning to the pictures, Table 1 reveals that less than half (46%) of all loan listings included a picture.

However, the market seems to value the pictures, as 64% of the funded listings contained a picture. There

is an incredible diversity of pictures on the Prosper site, ranging from earnest looking couples, to dogs

wearing antlers, to pictures of nature scenery, and the occasional bikini-clad young woman. Among listings

with pictures, 65% included one or more adults as the central focus of the picture, and 21% included both

adults and children. Another 10% were pictures of just children without adults. A sizeable (though smaller)

fraction of pictures contained no people, including 4% that were primarily of a building (e.g., a home or

storefront), 4% primarily picturing animals (e.g., pet dog), and 2% picturing an automobile.

For pictures that included adults, coders were instructed to code a number of perceived

characteristics. These include, gender, race, age, happiness, weight, and attractiveness. We also included

categories of secondary interest, such as whether the people were professionally dressed or displayed signs

of military involvement.20

The right-hand side of Table 1 gives summary statistics for the information coded from the pictures.

Looking first at gender, there is a rough balance between men and women in the genders displayed in the

loan listings. Of the pictures with people, pictures of single males make up 38% of the full sample and 40%

of the funded listings. The analogous figures for females are 35% and 31%, and for male-female couples are

20% and 22%. The coders also recorded the perceived race of the people pictured, using the primary

18 The one-line descriptions may be a more first-order influence on the lending decision than the longer descriptions. When prospective lenders browse loan listings, they first see a large page of listings (similar to a results page on Ebay), on which listings can be sorted or limited by a number of criteria. On this initial page, lenders see: a) the loan parameters (i.e., size, current interest rate, percent of the requested loan that has been funded, and the number of bids), b) credit grade and DTI, c) a picture (if provided by borrower), and d) the borrowers one-line loan description. Thus the picture and the one-line description are the information that lenders have when deciding which of the roughly 4,000 listings active at any one time to look at in detail. 19 For example, the correlation between the number of words in the longer text description and listings with a low credit grade of “HR” is -0.01, with white listings is 0.03, and with black listings is -0.03. 20 For each of these characteristics, the coding options included an unclear/uncertain category. Indicator variables for these unclear/uncertain categories are included throughout the analysis, but have very small cell counts, and to save space we drop them from our summary statistics and regression tables.

16

categories of white/Caucasian, black/African American, Hispanic/Latino, and Asian.21 The majority, 67%,

appear to be white, while 20% are coded as black, 3% as Hispanic, and 3% as Asian.22 Looking at the

listings that actually funded reveals that (unconditionally) minorities are much less likely than whites to

receive loans on Prosper – 83% of the funded listings with adult pictures were of apparently white

individuals. The patterns for age, weight, and the secondary characteristics are all sensible and reveal

relatively little difference between the full sample of listings and the listings that fund.

Comparing the distributions of these variables between the full sample of listings and the funded

listings suggests that the market: 1) favors pictures of whites over minorities by a significant margin, 2)

modestly favors pictures of men over women, of happy people over unhappy people, and thin people over

overweight people, and 3) does not react very strongly to the stated purpose of the loan. Of course, since

these characteristics may be highly correlated with other financial characteristics, the summary statistics

could be misleading.

IV. Empirical Results

Probability of Funding

In this section we investigate how the information contained in pictures and descriptions affects the

probability of funding holding all else equal. The summary statistics in Table 1 provided the first hint that

disparate treatment may exist in funding decisions. Figure 2 provides additional suggestive evidence. Figure

2a illustrates the funding rate by each credit grade by white and black borrowers. Two main findings can be

taken from this figure. High credit grade borrowers are more likely to be funded than low credit grade

borrowers, and whites are more likely to be funded than blacks at every credit grade. Figure 2b and 2c are

less conclusive, but suggest that females may be more likely to be funded than males (especially at lower

credit grades) and that older borrowers are less likely to be funded than younger borrowers.

21 These codings may not always agree with the race the borrower would list for him or herself if asked; however, it is the perception of race as conveyed through the pictures and not the actual race of borrowers that may affect lenders’ decisions. 22 Compared to statistics for the overall population from the 2000 Census -- White (73.9%), Black (12.2%), Hispanic (14.8%), and Asian (4.4%) – blacks are overrepresented in our sample, while whites and Hispanics are underrepresented.

17

As always, the challenge here is to overcome problems associated with omitted-variable bias so that our

estimates can reasonably be interpreted as the market response to the information provided by borrowers.

Fortunately, the Prosper data are ideally suited to this type of analysis. Unlike most other studies of credit

markets, the data available here contain all of the information about a listing that is seen by prospective

lenders. Of course, we are still challenged with the difficulties of using the available information correctly

and the usual problems that arise with the need to make functional-form assumptions. We address these

issues by using flexible functional forms on credit controls in our baseline specifications, and incorporating

numerous robustness checks, including a propensity-score analysis for our findings on race.

Our basic empirical strategy involves estimating the probability that a loan listing gets funded as a

function of the listing characteristics that are observed by the lenders. We use both linear probability

models, estimated via OLS, and Logit regressions. The basic linear regression framework is:

𝑌𝑖 = 𝛼 + 𝑋𝑖𝛽 + 𝑍𝑖𝜃 + 휀𝑖 ,

where Yi is an indicator variable for whether or not listing i was funded, Xi is a matrix of characteristics

coded from the pictures and one-line description of the purpose of each loan, and Zi is a matrix of other

characteristics of the listing and borrower, including credit controls and loan parameters. The regressions are

estimated over the full sample of 110,333 listings made during the one-year sample period. Because many

borrowers relist their requests when their listings expire without funding (generally with higher maximum

interest rates), we cluster at the borrower level to obtain standard errors.

Baseline Regression Estimates. Our baseline regression specification includes indicators for the

characteristics coded from pictures and text along with a large set of flexible controls for the other

parameters of the loan listing. These controls (i.e., Zi) include credit grade crossed with a cubic of the

maximum interest rate the borrower set, a cubic of the size of the requested loan, the duration of the loan

listing, the log of self-reported income, and a cubic of DTI. The other variables from a borrower’s credit

profile available to lenders are: number of current delinquencies, delinquencies in the last seven years, total

number of credit lines, total number of open credit lines, number of inquiries in the last six months,

revolving credit balance, and bank card utilization. These variables are included in the regressions in log

18

form.23 We also include dummy variables for homeownership status, occupation type, employment status,

whether the borrower was a member of a group, and the rating (one to five stars) of the group. Additionally,

we include variables created using our text analysis from the long-description: the log number of total

characters, a readability index (which uses word and sentence length), and the percent of words which are

misspelled. Finally, since this is an evolving market and one that can be affected by fluctuations in the

overall economy, we include month dummies to capture time effects unrelated to specific listing parameters.

The estimated coefficients on credit and loan-parameter controls (i.e., 𝜃 ) are sensible and unsurprising and

generally highly statistically significant. Because these variables enter the regression nonlinearly or with

interaction effects and due to space constraints, we do not report the coefficients here. However, later we

discuss a robustness table that shows estimates for some of these variables from a simpler linear

specification.

Table 2 shows the coefficient estimates for the variables we coded from the pictures and descriptions

(i.e., 𝛽 ). Columns (1) and (3) display the results using OLS and columns (2) and (4) display the Logit results

as the marginal effects of the variables on the probability of funding. For each of the categories listed in the

table, we have also listed the base-group on which the coefficient estimates are based. In order to use all of

the available data in our regressions, we included dummy variables to indicate when a listing had no picture

or a picture without people in it. The coefficients on these dummies are not reported in the table, since they

depend on the base-groups chosen for the race, gender, age, and other controls. However, in a similar

regression that includes the same credit controls, but codes only for whether or not a listing had a picture,

we find that listings without pictures are approximately 3 percentage points less likely to fund.

Consistent with the raw summary statistics, the largest effects of the picture characteristics are for race.

The OLS estimates imply that listings with a picture of an apparently black or African American person are

3.2 percentage points less likely to get funded than an equivalent listing with a picture of a white person.

Relative to the overall average funding rate of 9.3%, this is a 34% drop in the likelihood of funding. The

marginal effects from the Logit regression imply a slightly smaller but still economically meaningful

23 To avoid problems associated with ln(0), we added 1 to each variable before taking the log.

19

difference of 2.4 percentage points. Both estimates are statistically significant at the 1% level. Interestingly,

the negative effect of a black picture is approximately the same as that of displaying no picture at all.

After controlling for credit characteristics, the estimated effect of displaying a picture of a woman is the

reverse of what we saw in the summary statistics. In the raw summary statistics, women are less likely to

have their loan requests funded, but this is driven by the correlation between female pictures and credit

score. The estimated effects in Table 2 are positive, and in the Logit specification imply that all else equal

listings with a picture of a woman are 1.1 percentage points more likely to fund. This result is statistically

significant and approximately half the size of the estimated effect of a black photo.

The apparent age of the person in the picture is also an important predictor of successful funding.

Compared to the base group of 35-60 years old, those who appear younger than 35 have a predicted rate of

funding that is between 0.4 and 0.9 percentage points higher, while those who appear to be over 60 years

old are between 1.1 and 2.3 percentage points less likely to succeed in acquiring a loan. However, it is worth

noting that the elderly comprise only 2% of the pictures in the sample.

There are also some interesting results related to the perceived happiness, weight, and attractiveness of

individuals in their pictures, though the results are generally somewhat weaker. For instance, the OLS

estimates imply that listings of significantly overweight people are 1.6 percentage points less likely to fund,

which is statistically significant at the 5% level. However, the marginal effect in the Logit specification is

only -0.6 percentage points and is not statistically distinguishable from zero. The coefficients on our

measures of attractiveness imply directionally that more attractive people are more likely to have their loans

funded; however, the coefficient estimates are rather small and are not statistically significant.24 The

strongest effects from this set of characteristics are for perceived happiness. People who look unhappy are

between 1.6 (Logit) and 1.8 (OLS) percentage points less likely to have their loans funded. While these

24 In other specifications (not reported) we interact gender with this attractiveness measure to see whether there is an effect of pictures of especially attractive females. The estimates are in the direction of a positive interaction between female and attractiveness, but the magnitude is very small and statistically insignificant. We suspect that the inherent subjectivity of attractiveness and the coarseness of the measure we used may have introduced measurement error and subsequent attenuation bias in the attractiveness variable. Our results are directionally consistent with those of Ravina (2008), who conducted a more thorough coding of attractiveness using a smaller sample of Prosper loans and finds a strong positive effect of beauty on the likelihood of funding.

20

differences are statistically significant at the 10% level in both specifications, it is important to note that

unhappy people make up only 1% of all pictures.

Finally, we coded some secondary characteristics of pictures with adults, including whether the adult

had a child with them in the photo, whether the person was professionally dressed (e.g., wearing a tie), and

whether there were signs of military involvement (e.g., uniform). We find no significant effect of a child in

the picture or of professional dress on funding. While statistically insignificant in the OLS specification, in

the Logit specification military involvement increases the likelihood of funding by 2.5 percentage points.

The estimated effects of the coded loan purpose are generally weaker than those of the picture

characteristics, though there are some important and sensible patterns. The base-group for these purpose

dummies is the listings with no clear purpose that could be discerned from the one-line loan description.

Relative to that group, the loans listings that express interest in consolidating or paying down debt (usually

high-interest credit-card debt) are between 0.4 (Logit) and 0.5 (OLS) percentage points more likely to get

funded. Loans with most other purposes are less likely to fund, though many of the effects are not

statistically significant.

Robustness. Specification Checks: In Table 3 we begin to investigate the robustness of these results,

focusing on the estimated effects for race. The table reports marginal effects from the Logit regression for a

number of specifications. In the first column the regressors include only the gender and race characteristics

coded from the pictures without any credit or loan-parameter controls. They confirm the summary

statistics; blacks are 5 percentage points less likely to get funded than whites. The second column adds

dummies for the borrower’s credit grade, continuous linear measures of the maximum interest rate, DTI,

and requested loan size. Adding these controls brings the estimates much closer to the estimates reported in

Table 2, and highlights the important correlations that race has with credit measures; the estimated effect of

being black falls to -2.8 percentage points. This column also provides easy comparisons of the size of the

race effect. The marginal effect of being black (-2.8%) is somewhat less that the -4.1% effect of moving

from a credit score of above 760 (AA credit) to a credit-score range of 720-759 (A credit), and about one

and a half times as large as the effect of a one percentage-point change in the maximum interest rate.

21

Columns (3) through (6) of Table 3 add in interaction terms in the financial variables, additional credit

controls, the long-description text-analysis controls (e.g., percent of words misspelled), and time trends.

There is a slight drop in the race effect when additional credit controls are added, but otherwise the effect of

a black photo does not change meaningfully with these additional characteristics. Column (7) adds in the

other picture controls (e.g., professional dress) and column (8) adds in the loan purpose variables,

reproducing the regression from columns (2) and (4) in Table 2. Adding these other characteristics

strengthens the race effect slightly.

There are a few main takeaways from this robustness table. The first is that approximately half of the

disparity in loan funding between blacks and whites observed in the sample averages can be accounted for

by the different financial characteristics of black and white borrowers. It is also important to note that once

basic credit controls are included in the regressions, the estimated effects on race are quite stable across

different specifications.

Another potential worry is that despite the controls we use, our coding procedure may fail to fully

capture impressions about educational status or income that lenders infer from the pictures and descriptions

that borrowers provide, and that these inferences may be correlated with race. Adding in controls for

measures of the borrower’s self-reported income and occupation, however, barely affect the estimates for

blacks versus whites. Thus to the extent that inferred education and income are correlated with stated

income and occupation, this would suggest that differential inferences about education or income stemming

from listing features that are observable to lenders (but difficult for the econometrician to incorporate) are

not of great concern.

Propensity-score matching: Despite these robustness checks, one might still worry about whether whites

and blacks are similar enough for us to find ceteris paribus effects of race. For instance, it could be the case

that blacks anticipate discrimination and set systematically higher maximum interest rates than their white

counterparts making it difficult to compare blacks and whites that match well on observables. We find in a

regression of interest rates on listing variables that black borrowers do not set systematically different

interest rates than whites and if anything set slightly lower rates all else equal. Nonetheless there could be

other important variables for which it is hard to find similar blacks and whites. Therefore, as a further

22

robustness check we employ a propensity-score matching estimation (Rosenbaum and Rubin, 1983; Dehejia

and Wahba, 1999).25 Although there is no traditional “treatment” variable in this setting, for the purposes of

this robustness check, we focus solely on the racial differences and treat listings with black photos as the

treated sample and listings with white photos as the untreated sample.26

In the first step of this process we estimate the probability that a listing includes a black photo (as

opposed to a white photo) using a Logit regression of a black-picture dummy on all other listing variables.

For this regression we use a very flexible specification that includes multiple polynomials of important

variables (e.g., borrower’s max rate and dti) as well as interactions of credit grade with other controls and

borrower’s maximum interest rate with other controls.27 The predicted values form the propensity score for

being black. Appendix figure A1.a shows histograms of the estimated propensity-score distribution for

blacks and whites. Although blacks have higher propensity scores on average, there is substantial overlap in

the distribution, with only 98 blacks having estimated propensity scores above the maximum propensity

score among whites. In the matching procedure we drop these black listings that are off the support.

In the next step of the estimation we match each of the black listings with the white listing that has the

most similar estimated propensity score.28 Appendix figure A1.b. shows that this nearest-neighbor matching

procedure is very effective at reducing the differences in covariates between the black and white listings.

The figure displays the standardized bias between black and white listings (Rosenbaum and Rubin, 1985) for

a number of key listing variables both before and after matching.29 Before matching, a number of covariates

have large bias measures; for example, in the unmatched sample blacks are much more likely to be females

and have the high-risk credit grade. After matching, however, virtually all of the covariates have

standardized bias under 5% and most under 2%, suggesting that the nearest-neighbor matching procedure is

25 The propensity-score approach only allows us to focus on one treatment effect (e.g., black). Thus, we continue to report OLS and Logit estimates in our main tables since we are interested not just in the coefficients for race, but for the coefficients on a wider set of variables. 26 Listings without pictures or with pictures that do not convey race of either black or white are dropped from this analysis. Restricting the sample in this way leaves us with 29,225 listing observations, of which 6,515 (22%) are black. 27 All variables with the exception of the state dummies are included in these interactions. Interactions with state dummies create too many variables, many with small cell counts, and create problems with convergence in the logit regression. 28 All of the nearest-neighbor matches for the blacks used in the matching procedure are within a caliper of 0.01 around the estimated propensity score. 29 The standardized bias for a covariate is defined as the difference in sample means of the variable (black – white) as a percentage of the square root of the average of sample variances for the covariate in the two groups.

23

effective at creating balanced samples.30 Using this matched sample we calculate the average treatment on

the treated and find a difference in the likelihood of funding between blacks and whites of 3.2 percentage

points (p < .01), exactly the same as in the OLS regressions in Table 2.31 The consistency of the race

estimates using propensity-score matching suggests that the results are not the artifact of a flawed regression

specification.

Alternative sample cuts: In Table 4 we investigate the race results under a number of different cuts of the

data. Each cut uses the baseline Logit specification from Table 2 and reports marginal effects. Cutting by

credit grade reveals that across all credit grades there is a significant negative response to black pictures.

The percentage point difference in the likelihood of funding between blacks and whites is actually higher for

better credit grades: blacks are between 4 and 6 percentage points less likely to be funded amongst

borrowers with credit scores above 640 (grades of C and above), compared to a 3.3 percentage point

difference for D&E credit (560 – 640) and a 1.3 percentage point difference for the high-risk borrowers

(520 – 560). Comparing these differences to the mean probability of funding for the different groups,

however, reveals that the likelihood of funding is 37% lower for blacks in the high-risk category versus

12.2% for blacks in the highest credit grades.

The second cut we investigate splits the one-year sample in half and contrasts results estimated over

listings in the first six months of the sample versus those in the second six months. None of the results are

meaningfully different between these samples. Although the market itself is evolving rapidly, the market

response to information contained in pictures and text has remained relatively stable.

For the third cut in Table 4, we divide the sample into quartiles of self-reported income. The negative

marginal effect of a black picture versus a white picture is slightly larger for higher income quartiles –

ranging from -2.1 percentage points for the lowest income quartile to -3.4 percentage points for the highest

income quartile. Of course, these income quartiles have different mean rates of funding, and thus in

percentage terms the negative effect of a black picture is quite a bit larger in the lowest quartiles.

30

The average bias across all covariates falls from 8.4% in the unmatched sample to 1.0% in the matched sample. 31 An alternative matching procedure using radius matching with a caliper of 0.05 gives an average treatment on the treated of -2.8 percentage points.

24

The final cut in the Table 4 investigates whether the race and gender effects vary depending on the

borrower’s stated occupation. We split the sample based on occupations that are likely to require a college

degree versus those that do not. The negative marginal effect of a black picture is slightly more than a

percentage point larger for those with high education jobs (-3.3% to -1.9%). When compared to the

funding base rates of the two groups, however, the marginal effects are quite similar in percent terms. The

fact that the results for blacks are not strongly related to these occupation cuts, again suggests that any

failure on our part to fully capture inferences that lenders can make about educational attainment of the

borrowers based on observables is unlikely to explain the race results.

One final note on the robustness of our estimates of the probability of funding is in order. Lenders

have the option of creating settings that automatically bid on loans based on lender-chosen criteria of credit

score, DTI, and the like. We are not able to ascertain how many lenders use this option, but if all lenders

exclusively used this process, we would not find any effect on the picture or text characteristics. Hence our

results may underestimate the market response that would be observed in a market without automatic

bidding. The results also highlight that market participants do in fact react to the non-financial information

and that many forgo the option to bid on loans without reviewing the listing in detail.

Final Interest Rate on Funded Loans

The differences in the likelihood of funding translate into different final interest rates conditional on a

loan getting funded. Table 5 presents the results from an OLS regression of the final interest rate of a

funded loan on the borrower and listing characteristics used in the baseline specification (Table 2), excluding

the maximum interest rate the borrower set. The estimates are in the directions one would expect based on

the estimates of the probability of loan funding. The first column of the table is estimated over all 10,207

loans made in the Prosper market during our sample year. All else equal, a funded listing with a picture of a

black borrower ends up with an interest rate that is 60 basis points higher than an equivalent listing for a

white borrower. Single females have rates that are 40 basis points lower than males. The results for age and

happiness are much smaller and not statistically distinguishable from zero. The very unattractive end up

with rates that are 60 basis points higher than their average-looking counterparts. The effects of the stated

25

loan purposes are also sensible given the results above. For instance, those expressing a desire to

consolidate credit-card debt obtain loans with interest rates that are 20 basis points lower than their

counterparts who express a need for a business loan.

These estimates are consistent with the predictions of the idea that the different reservation rates

lenders set for loans from otherwise similar “majority” and “minority” borrowers would lead to different

interest rates on funded loans for the groups. However, there is a potential problem with interpreting these

interest-rate results in that way. Borrowers may elect to forgo the “bid-down” process and receive their loan

funds at the maximum interest rate they set as soon the loan becomes fully funded. The worry here is that

if, for example, black borrowers were more likely to use this feature and occasionally set maximum interest

rates that were highly attractive to lenders, it might result in higher interest rates for funded black loans than

similar funded white loans, even if the reservation rates of the lenders were the same for the two groups.

To address this concern, column (2) of Table 5 restricts the analysis to the 6,419 funded loans that used the

“open funding” option that allows interest rates to be bid down to the reservation rate of the marginal

lender. The results are quite similar, and in fact the effect of a black loan increases from 60 basis points in

column (1) to 80 basis points for the loans that allow bid down.

Net Return on Funded Loans

The preceding analysis reveals that the market discriminates based on information contained in pictures

and text and that this discrimination leads to disparities in interest rates on funded loans. Here we ask

whether the discrimination we observe in the Prosper.com market is efficient for the lenders.

Loan-Performance Data.32 Prosper provides performance data on all loans that have been made in

the marketplace. The analysis here is based on the available performance data as of December, 2007, at

which point the loans made during our sample year ranged in age from 7 months to 19 months. Prosper

provides information on payment status of each loan showing whether the loan was current, paid off, 1

month late, 2 months late, 3 months late, 4+ months late, and officially defaulted. Table 6 shows summary

32 Disclaimer: None of the loans made on Prosper.com have reached full maturity. Because of this, all estimates of loan profitability in this marketplace are only valid subject to the assumptions discussed.

26

statistics for this performance data, combining defaulted loans with those that are 4+ months late (which

exceeds the usual standards for considering a loan in default). Among all loans made during the sample

year, as of December 2007, 78% had been paid off or were in good standing. Approximately 2% fell in each

of the categories, 1 month late, 2 months late, and 3 months late. A sizeable fraction (17%) of all loans was

4 months late or more.

The table breaks down the performance data by the age of the loan. Naturally, the number of loans in

good standing is higher for the more recent loans. For instance, only 2% of the loans made in May, 2007

were 4 months late or more in payments as of December, 2007, which is sensible when one considers that

these borrowers would have had to stop paying by the third month of their loan to fall into this category.

The payment characteristics are rather stable, however, for loans that are at least 13 months old, suggesting

that most default in the Prosper market may occur during the first year of the loan. Somewhere between

71% and 74% of loans were in good standing after 13 months, while 20-25% were at least 4 months late in

making payments.

Table 6 continues by indicating loan performance information by race. There are large differences in

default rates across racial groups. Most notably, 29% of loans made to black individuals are 4+ months late.

In comparison, only 14-15% of loans made to white or Asian borrowers are 4+ months late. Hispanic

default rates fall in between these groups with 21% of loans that are 4 or more months late.Once again,

however, given the correlations that exist between these groups and other variables (e.g. credit grades), the

summary statistics do not provide conclusive evidence that these groups have higher default rates

controlling for all of the other information available to lenders.

Hazard Model of Default. In order to formally test whether there exist ceteris paribus differences in

default rates across gender, race, and other groups, we employ a simple hazard model where default is

considered to be a nonnegative random variable. We estimate the hazard function, 𝜆(𝑡), as defined in the

analysis of Cox’s proportional hazard model (Cox, 1972). 𝜆 𝑡 measures the instantaneous failure rate at

time t given that the individual survives until time t. In our model, a “failure” is a loan that goes into

default. For this model, we define a loan as entering default when the borrower misses three consecutive

27

pay cycles (a common assumption in the literature on loan repayment). In this model the baseline hazard

rate, 𝜆0(𝑡), remains unspecified and through the exponential link function, the same covariates 𝑋𝑖 and 𝑍𝑖

that are used in our baseline regressions in Table 2 act multiplicatively on the hazard rate.

The hazard-model estimation results are presented in Column (1) of Table 7. Once again, the largest

and most significant effects that we find in the estimation of the hazard model involve the race variables.

Blacks are approximately 36% more likely to default on their loans than are whites with similar

characteristics. The summary statistics indicated that blacks were twice as likely to default as whites. While

the estimate on the black coefficient is smaller after controlling for credit and other variables, it is still

statistically significant and obviously economically large. While not statistically significant, Asians and

Hispanics are estimated to be 24% less likely and 10% more likely to default than whites, respectively.

Few of the coefficients on the other picture characteristics are statistically significant. However, the

direction of the effects is interesting. The parameter estimates suggest, for instance, that women are 14%

more likely to default than men. The difference across age groups is essentially zero. Borrowers that the

coders recorded as appearing unhappy are estimated to be 42% more likely to default. The results of being

somewhat or very overweight relative to being thin are mixed. Borrowers coded as being very unattractive

are estimated to be 32% more likely to default, though not significant and borrowers that indicated signs of

military involvement are estimated to be 49% more likely to default.

Estimates of Net Return. The default data alone do not tell us about discrimination in this market.

Differences in default rates are necessary for the earlier results to be explained by accurate statistical

discrimination, yet they are not sufficient. In order to answer this question, we need to combine the default

rate data with the interest rates that borrowers actually paid. We begin with a simple graph. Figure 3

presents the fraction of loans defaulted versus the final interest rate on the loan using linear smoothing.33

It is clear from the graph that at each interest rate, the proportion of black loans defaulting is higher than

the proportion of white loans defaulting. Figure 4 adds to the evidence presented in Figure 3 by illustrating

the dynamics of loan performance. We begin by calculating the returns lenders see as an average annual

33 This graph makes use of the lowess command in STATA. For this Figure, we define a loan to be in default if the borrower has missed three or more consecutive pay cycles.

28

percentage rate (APR) across black and white loans for each month as the loans age. Under the assumption

that borrowers who are current or less than two months behind will continue to make payments until loan

maturity and that borrowers that are three or more months behind are in default, we graphically

demonstrate how the average APR by race declines over the maturation of loans as defaults begin occurring.

Figure 4a and 4b illustrate this by looking at loans for which we have at least 12 and 15 months of loan-

performance data, respectively. As can be seen in each of these figures, black loans have a higher APR at

the beginning (due to the fact that they are required to pay higher interest rates on average). As loans

mature, however, the higher default rate on black loans causes the net return on these loans to fall below the

net return on white loans. In fact, the black default rates are such that by 4 months the net return on black

loans is lower than the net return on white loans, at 9 months the net return on black loans is negative, and

at 12 months the average APR on black loans is approximately -5% relative to a 5% APR for the average

white loan.

While Figures 3 and 4 provide a nice visual representation of loan repayment by race, it does not allow

us to estimate the difference in net return for race while controlling for credit grade and other important

variables. In order to do this more rigorous analysis, we begin by calculating the net return over a three-year

period on a dollar invested.34 The calculation uses the monthly payment on the loan, and thereby

incorporates the interest rate on the loan. We consider three different measures of net return, based on

different assumptions about the future repayment of loans. Each measure assumes that any loan that is in

good standing (current or paid off), in December of 2007 will continue to be paid off throughout the

remainder of the three-year loan period. This assumption is obviously generous, as some of the loans in

good standing will default in the future. Furthermore, assuming that the loans that are paid off earn the full

three-year return, is equivalent to assuming that lenders who are paid early can costlessly find another loan

with the same terms. The differences between the return measures come from different assumptions about

the repayment stream for loans that were late as of December 2007. Our first net return variable, Return

34 In Figure 4, we use APR as the relevant statistic for evaluating the net return on a loan. We are unable to use this measure in the regression analysis due to the fact that APR is undefined when a borrower does not make any payments on a loan. Thus, while APR could be used when we were looking at averages across a group, we employ the 3-year net return on a dollar as the relevant statistic for the individual-level regressions.

29

Type I, is the most pessimistic about future loan performance; we assume that any loan that is not in good

standing (1 month or more late) in December of 2007 will not produce any future payments. For Return

Type II we assume that a loan that is only 1 month late will eventually pay in full, and for Return Type III

that a loan that is 2 or fewer months late will eventually pay in full. Hence, these return types are

increasingly optimistic in that they assume that loans that are in good standing as well as loans that are only

1-3 months late will all be paid in full.

Columns (2)-(4) of Table 7 show the results of OLS regressions of net return on the baseline covariates

that have been used throughout for the full set of funded loans. The top row provides the mean of the net-

return variable across all funded loans for each return type. The average 3-year return on a dollar lies

between 1.047 (Return Type I) and 1.084 (Return Type III). Translating these three-year returns into annual

percentage rates yields a net APR range of 3.1% to 5.3%.35

Turning to the regression estimates, as before, we present the estimated coefficients and clustered

standard errors for the various picture characteristics. Across the different return types, the only variable

that is consistently statistically significant is the black indicator variable. The estimates for the full sample of

funded loans suggests that the average net return on investing in a loan from a black borrower is 8.2 to 8.6

percentage points lower over a three-year period than investing in a loan from an otherwise similar white

borrower. This result implies that the increased propensity of default for black loans was not fully offset by

the one percentage point increase in interest rates that black individuals paid. While not significant, the net

return on Hispanic loans is estimated to be 3.1 to 5.5 percentage points less than whites while the net return

on loans given to Asian borrowers is estimated to be 1.5 to 1.7 percentage points higher than whites.

The estimated coefficients on the other variables are not consistently statistically significant, but the

direction of the coefficients may be of interest nonetheless. The estimated return on loans to single females

is approximately 2 percentage points less than for single males. The return for borrowers coded as being

unhappy, older, unattractive, professionally dressed, involved in the military, or with a child is less than their

35 It is worth noting that these low average returns may not bode well for the long-run viability of the Prosper model. In fact, our return measures are especially generous, because many of the loans in the sample have been out for less than a year.

30

counterparts. Conversely, the return on borrowers coded as very overweight is higher than their

counterpart.

Columns (5) through (7) of Table 7 present the same analysis estimated over the sample of loans that

had “open funding”, indicating that the listing remained open after reaching funding level, which allows

lenders to continue to bid down the interest rate set by the borrower. Although slightly smaller, these

estimates show very similar differences in net returns between whites and blacks, with blacks having net

returns that are 7.3 to 7.5 percentage points lower over the course of three years.

Overall, the net-return results do not appear to differ significantly based on the different return-type

definitions. Nonetheless, it is worth thinking about how our estimates might be affected by the assumptions

that we use to generate our estimates of net return. Specifically, is it possible that the strong negative return

effect that we find for blacks could change once all 3 years of return data are available? The estimate on the

black coefficient would be attenuated if the default rate of whites becomes significantly larger than the

default rate of blacks after December of 2007. Given the available data, this possibility can never be ruled

out. However, the available data do not suggest that this is the case. In fact, if anything, loans given to

black borrowers are defaulting at a higher rate after the loan has matured for a year or longer; indicating that

we might be underestimating the overall effect of race on net returns. Furthermore, defaults that occur later

are less costly to lenders because more of the principle has been repaid. Hence, the difference in default

rates as loans mature would have to change dramatically to erase the coefficient that we estimate for the

black indicator variable.36

Another potential concern with the results – one that arises in many studies of default – is that an

unanticipated economic shock that hit blacks harder than whites could affect the results. That is, lenders

might have charged interest rates for each group that equalized net returns ex ante but turned out to wrong

ex post. While it is difficult to rule out this possibility, two different cuts of the data suggest that differential

economic shocks are unlikely to explain the differences in average net returns. First, we divide the sample

into those funded during the first 6 months and second 6 months of our one-year sample. If unanticipated

36

As an example of how dramatic the change in trends in default rates would have to be, if we assume that the total percentage of people defaulting each month continues to be the same over the life of the loans, we estimate that whites would have to immediately begin defaulting at three times the rate of blacks in order to close the gap in net returns.

31

economic shocks are the source of the net-return differences, we would not expect the results to be the

same across periods. Although this cut causes us to lose power and is complicated by the different time

horizons for default in the split samples, we find no evidence of differences in the net return results for the

two samples. As a second approach to addressing the possibility of differential economic shocks, we exploit

the state identifier on listings to control for changes in the economic environment that occur during the

repayment process. To do this, we obtain 2006 and 2007 unemployment rates by state and race from the

Local Area Unemployment Statistics (LAUS) prepared by the BLS. We re-estimate Columns (2) – (4) of

Table 7 including an interaction term between the race variable and the difference in the unemployment-rate

change between 2006 and 2007 between blacks and whites. After controlling for these changes in economic

conditions that may differential affect racial groups, we find estimates for each return type ( -.074, -.082, and

-.084) that continue to be statistically significant (p < .01 for each return type) and are similar to the

estimates found in Table 7.

V. Discussion and Conclusions

We have shown that the characteristics borrowers display through their pictures and descriptions

strongly affect their access to credit in the Prosper market. Specifically, we find significant discrimination

against listings without a picture and listings with pictures of blacks, older individuals, and people who

appear unhappy. In contrast, there is discrimination in favor of listings with pictures of women and pictures

that show signs of military involvement.

If this discrimination was solely the result of costly taste-based preferences, we would expect a negative

correlation between the discrimination in funding and subsequent net returns on funded loans. Accurate

statistical discrimination, on the other hand, should result in no significant differences in net returns. The

results for black loans run counter to the predictions of both taste-based animus against blacks and accurate

statistical discrimination. Despite the fact that blacks are less likely to have their loans funded, the return

results would suggest some form of prejudice in favor of blacks. That is, although blacks pay higher interest

32

rates, those rates are not high enough to account for the higher probability of default that we find for black

loans after controlling for other observable characteristics.

How can we interpret these findings? First, we note that clearly skin color is not a causal factor in loan

default. Higher default rates for blacks must stem from some difference in the background and financial

characteristics of these borrowers that is not fully reflected in the standard financial measures (e.g., credit

score, DTI) that lenders can observe. There are a number of well-known candidates for important

characteristics that are not (perfectly) observed by lenders: for example, income disparities (perhaps

stemming from labor-market discrimination), education differentials, and more limited access to financial

support from family and friends.37 Whatever the relevant differences, the discrimination in the decision to

lend suggests that the market understands the direction of this correlation between unobservables and the

race in a borrower’s picture. Yet the interest-rate penalty these borrowers pay in the market is not enough

to account for their higher rate of default.

One explanation for these results could be a combination of accurate statistical discrimination against

blacks (based on their greater likelihood of default conditional on other observables) that is partially offset

by taste-based discrimination against whites (in favor of blacks). That is, the market might accurately assess

the default probability of loans on average, but lenders might have a taste for lending to blacks over whites

and that makes them willing to accept a lower return from these loans. If true, this finding in peer-to-peer

lending is quite novel, as to our knowledge the previous literature on discrimination has never found

evidence of taste-based discrimination that favored blacks over whites.

Another potential explanation of the results is that while lenders understand the direction of the

correlations between race and relevant unobservable characteristics, they fail to appreciate the strength of

these correlations or the importance of unobservable characteristics on default. Although it seems

intuitively plausible, this explanation implies that biased beliefs exist at the market level. We might expect

these types of mistaken beliefs to be driven out of the market in a long-run equilibrium, and perhaps the

37 One example that we can test comes from the credit score. Prosper only provides credit-score grades to the lenders and it is possible that within a credit grade credit score is correlated with being black. Although lenders do not see the credit score, Prosper provided us with this information. Perhaps surprisingly, however, we find that being black is not correlated with credit scores within a credit grade when other variables are controlled for, and thus cannot explain the net-return differences.

33

evidence here is simply consistent with a market that is still evolving. Yet it is important to note that in this

case, we are finding that a market with an efficient auction mechanism, real stakes, and large amounts of

available data on performance is still not at its long-run equilibrium two-years out. In fact our splits of the

data reveal little change in the discrimination against blacks between the first half and second half of our

sample.

The results here also have implications for the broader literature on assessing theories of discrimination.

Had we found that blacks have higher net returns, it would have been natural to conclude that the evidence

was consistent with taste-based preferences against blacks. Having instead found the opposite, however, we

are faced with the somewhat awkward conclusion that the evidence is consistent with partial taste-based

discrimination in favor of blacks over whites. The alternative, which seems somewhat natural in this setting,

is to conclude that decision-makers have inaccurate beliefs. The problem, of course, is that once one allows

for the possibility of inaccurate beliefs, results from other studies that find evidence of taste-based or

accurate statistical discrimination come into question. Thus, the results from this study suggest caution

when interpreting evidence in favor of one theory of discrimination versus another.

The findings in this paper also highlight the importance of attempting to assess the efficiency of

discrimination before reaching conclusions about sources of discrimination. We find racial discrimination in

lending decisions despite the wealth of credit controls available to lenders, and it might be natural to

conclude that such evidence is suggestive of taste-based animus against blacks. Yet the data tell a very

different story that suggests that this peer-to-peer lending market actually treats the races more equally than

would be expected in a market with accurate statistical discrimination.

34

References

Aigner, D.J. and G. Cain. “Statistical Theories of Discrimination in Labor Markets.” Industrial and

Labor Relations Review,” 1977, 30, pp. 1975-1987.

Altonji, J.G. and R.M. Blank. “Race and Gender in the Labor Market,” in: O. Ashenfelter & D. Card,

eds., Handbook of Labor Economics, edition 1, volume 3, chapter 48, pp. 3143-3259.

Altonji, J.G. and C.R. Pierret. “Employer Learning and Statistical Discrimination,” Quarterly Journal of

Economics, 2001, 116(1), pp. 313-350.

Antonovics, K. and B.G. Knight. “A New Look at Racial Profiling: Evidence from the Boston Police

Department,” 2004, NBER Working Paper #10634.

Arrow, K. “The Theory of Discrimination,” in O. Ashenfelter and A. Rees, eds., Discrimination in

Labor Markets. Princeton: Princeton University Press, 1973.

Ayres, I., and P. Siegelman. “Race and Gender Discrimination in Bargaining,” American Economic

Review, 1995, 85(1), pp. 304-321.

Blank, R.M., Dabady, M., and C.F. Citro (eds). Measuring Racial Discrimination. Washington, DC:

National Academies Press, 2004.

Becker, G. The Economics of Discrimination. Chicago: University of Chicago Press, 1957.

Bertrand, M. and S. Mullainathan. “Are Emily and Brendan More Employable than Latoya and Tyrone?

Evidence on Racial Discrimination in the Labor Market from a Large Randomized Experiment,”

American Economic Review, 2004, 94(4), pp 991-1013.

Charles, K and J. Guryan. “Prejudice and the Economics of Discrimination,” 2007, NBER Working

Paper #13661.

Cox, D.R. “Regression Models and Life Tables (with Discussion),” Journal of the Royal Statistical

Society, Series B, 1972, 34, 187-220.

Dehejia, R.H. and S. Wahba. “Causal Effects in Nonexperimental Studies: Reevaluating the Evaluation

of Training Programs,” Journal of the American Statistical Association, 1999, 94:448, 1053-1062.

Freedman, S. and G. Z. Jin. “Dynamic Learning and Selection: the Early Years of Prosper.com,” 2008,

Mimeo.

Galster, G. “Comparing Loan Performance Between Races as a Test for Discrimination,” Cityscape,

1996, 2(1), pp.33-39.

Heckman, J.J. “Detecting Discrimination,” Journal of Economic Perspectives, 1998, 12(1), pp. 101-116.

35

Imbens, G. “Nonparametric Estimation of Average Treatment Effects Under Exogeneity: A Review,”

Review of Economics and Statistics, 2004, 86(4), pp. 4-29.

Knowles, J., Persico, N., and P. Todd. “Racial Bias in Motor Vehicle Searches: Theory and Evidence,”

Journal of Political Economy, 2001, 109(1), pp. 203-232.

Levitt, S. “Testing Theories of Discrimination: Evidence from Weakest Link,” Journal of Law and

Economics, 2004, 47(2), pp. 431.

Lundberg, S.J. and R. Starz. “Private Discrimination and Social Intervention in Competitive Labor

Markets.” American Economic Review, 1983, 73, 340-347.

Munnell, A.H., Tootell, G.M., Browne, L.E., and J. McEneaney. “Mortgage Lending in Boston:

Interpreting HMDA Data,” American Economic Review, 1996, 86(1), pp. 25-53.

Phelps, E. “The Statistical Theory of Racism and Sexism,” American Economic Review, 1972, 62(4), pp.

659-61.

Quigley, J.M. ”Mortgage Performance and Housing Market Discrimination,” Cityscape, 1996, 2(1), pp.

59-64.

Ravina, E. “Love & Loans: The Effect of Beauty and Personal Characteristics in Credit Markets,”

Mimeo, 2008.

Rosenbaum, P. and D. Rubin. “The Central Role of the Propensity Score in Observational Studies for

Causal Effects,” Biometrika, 1983, 70, 41-50.

Rosenbaum, P. and D. Rubin. “Constructing a Control Group Using Multivariate Matched Sampling

Methods that Incorporate the Propensity Score,” The American Statistician, 1985, 39, 33-38.

Ross, S.L. “Flaws in the Use of Loan Defaults to Test for Mortgage Lending Discrimination,”

Cityscape, 1996, 2(1), pp. 41-48.

Ross, S.L. “Mortgage Lending Discrimination and Racial Differences in Loan Default,” Journal of

Housing Research, 1996, 7:1, 117-126.

Ross, S.L. “Mortgage Lending Discrimination and Racial Differences in Loan Default: A Simulation

Approach,” Journal of Housing Research, 1997, 8:2, 277-297.

Ross, S.L. and J. Yinger. The Color of Credit: Mortgage Discrimination, Research Methodology, and

Fair-Lending Enforcement. Cambridge: MIT Press, 2002.

Tootell, G.M. “Redlining in Boston: Do Mortgage Lenders Discriminate Against Neighborhoods?”

Quarterly Journal of Economics, 1996, 111(4), pp. 1049-79.

36

Turner, M.A., Struyk, R., and J. Yinger. Housing Discrimination in America: Summary of Findings

from the Housing Discrimination Study. Washington, DC: Urban Institute Press, 1991.

Turner, M.A., Ross, S.L, Galster, G.C., and J. Yinger. Discrimination in Metropolitan Housing Markets:

National Results from Phase 1 of HDS2000. Washington DC: U.S. Department of Housing and Urban

Development, 2002.

Yinger, J. “Why Default Rates Cannot Shed Light on Mortgage Discrimination,” Cityscape, 1996, 2(1),

pp. 25-31.

Notes: This Figure illustrates monthly counts of the total number of listings and the total number of listings that

were eventually funded on Prosper.com since the company went public in February 2006. The loan listings that

we analyze in this study come from the 1-year period between June 2006 and May 2007. These dates are

indicated by the green vertical lines.

0

2000

4000

6000

8000

10000

12000

14000

2006m2 2006m6 2006m10 2007m2 2007m6

Nu

mb

er

of

List

ings

Date

Figure 1. Total Listings Across TimeBy Funded Status

Non-Funded Listings Funded Listings

Figure 2. Fraction of Listings Funded                       by Credit Grade and Demographics

0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

A_AA B  C D E HR

Fraction

 of Loa

n Listings Fun

ded

Credit Grade

Figure 2a.  Race

White Black

0 2

0.25

0.3

0.35

0.4

0.45

n Listings Fun

ded 

Figure 2b.  Gender

Notes:  Figure 2a illustrates the fraction of loan listings that funded for each credit grade by race.  The sample includes all loans between June 2006 and May 2007 that posted pictures where the race of the individual/s was discernable.  Credit grade bins are related to credit scores in the following manner:  A_AA (720 and up), B (680‐719), C (640‐679), D (600‐639), E (560‐599), and HR (520‐559).  Figure 2b uses data from loans during the same time period for which a picture was posted of a single adult male or a single adult female.  Figure 2c uses data from loans during the same time period for which a picture was posted of an adult/s where the age of the adult/s was judged to be "young" (less than 35), "middle" (35‐60), or "old" (more than 60).

0

0.05

0.1

0.15

0.2

A_AA B C D E HR

Fraction

 of Loa

Credit Grade

Male Female

0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

A_AA B C D E HR

Fraction

 of Loa

n Listings Fun

ded 

Credit Grade

Figure 2c.  Age

Young Middle Old

Notes: Figure 3 illustrates the relationship between the final interest rate on a loan and the fraction of loans with

that interest rate that have defaulted. Here we define default as a loan that is deliquent 3 months or more. This

relationship is illustrated separately for funded loans of borrowers that listed a picture of a black individual/s and

borrowers that listed a picture of a white individual/s. The lines are smoothed using a locally weighted

estimation (lowess) with a bandwith of 0.3.

0

0.1

0.2

0.3

0.4

0.5

0.6

0.1 0.12 0.14 0.16 0.18 0.2 0.22 0.24 0.26 0.28

Frac

tio

n o

f Lo

ans

Def

ault

ed

Final Loan Interest Rate

Figure 3. Final Interest Rate and DefaultBy Race

White Black

Figure 4. Average APR Adjusted Across Life of the

Loans by Race

Notes: This figure illustrates the average APR over the maturation of a loan for funds invested in

loans whose listings included a picture of a white individual/s or a black individual/s. The APR is

calculated at each month by assuming that the loan will be fully repaid if the loan has not gone into

default. We define a loan as goint into default if it is 3 or more months overdue. Thus the first data

point (that comes prior to the first payment), assumes that all loans will be paid in full and simply

illustrates the average interest rate paid by black and white borrowers. Each subsequent data point

is adjusted given the number of loans being defaulted by each group. Panel A illustrates the

dynamic APR for all loans for which we have at least 12 months of loan performance data. Panel B

illustrates the dynamic APR for all loans for which we have at least 15 months of loan performance

data. By restricting each panel to loans for which we have data over the entire x-axis time period,

we are able to graph this relationship without any attrition (each of the points on a line is reflected

by the exact same loans).

-0.05

0

0.05

0.1

0.15

0.2

0.25

0 1 2 3 4 5 6 7 8 9

AP

R

Months Since Loan Inception

Figure 4a. Tracking over 9 Months

White Black

-0.1

-0.05

0

0.05

0.1

0.15

0.2

0.25

0 1 2 3 4 5 6 7 8 9 10 11 12

AP

R

Months Since Loan Inception

Figure 4b. Tracking over 12 Months

White Black

Variables

All

Listings

Funded

Listings Variables

All

Listings

Funded

Listings

Credit Grade

AA (760-800) 0.03 0.10 Adult/Adults 0.65 0.67

A (720-760) 0.03 0.09 Just Children 0.10 0.07

B (680-720) 0.04 0.11 Buildings 0.04 0.05

C (640-680) 0.07 0.16 Animals 0.04 0.04

D (600-640) 0.11 0.17 Automobiles 0.02 0.02

E (560-600) 0.18 0.17

HR (520-560) 0.54 0.20

NC 0.01 0.01

Single Male 0.38 0.40

Loan Information Single Female 0.35 0.31

$ Requested 7,154 5,930 Couple 0.20 0.22

Borrower's Max Rate 0.17 0.20 Group 0.07 0.07

Final Rate NA 0.18

Fraction Funded 0.12 1.00 Race

Closed Auction Loans 0.43 0.37 White 0.67 0.83

Black 0.20 0.11

Other Information Provided Asian 0.03 0.03

Debt to Income Ratio 0.63 0.39 Hispanic 0.03 0.02

Group Member 0.51 0.69

Owns a Home 0.27 0.39 Age

Provided a Picture 0.46 0.64 Less than 35 yrs 0.53 0.54

35-60 yrs 0.41 0.41

Information Coded From More than 60 yrs 0.02 0.02

Descriptions:

Purpose of Loan Happiness

Consolidate or Pay Debt 0.30 0.33 Happy 0.74 0.77

Business/Entrepreneurship 0.10 0.10 Neutral 0.23 0.21

Pay Bills 0.04 0.02 Unhappy 0.01 0.01

Education Expenses 0.03 0.03

Medical/Funeral Expenses 0.03 0.02 Weight

Home Repairs 0.02 0.03 Not Overweight 0.73 0.75

Auto Purchase 0.02 0.02 Somewhat Overweight 0.20 0.18

Home/Land Purchase 0.02 0.02 Very Overweight 0.03 0.02

Auto Repairs 0.01 0.01

Luxury Item Purchase 0.01 0.01 Attractiveness

Wedding 0.01 0.01 Very Attractive 0.05 0.06

Reinvest in Prosper 0.01 0.02 Average 0.91 0.91

Taxes 0.01 0.01 Very Unattractive 0.03 0.02

Vacation or Trip 0.01 0.01

Multiple of Above Reasons 0.06 0.05 Other

Unclear/Other 0.34 0.33 Profesionally Dressed 0.13 0.14

Child Also in Picture 0.21 0.21

Signs of Military Involvement 0.02 0.02

Observations 110,333 10,207 50,820 6,571

Notes: This Table presents summary statistics for the 110,333 loan listings posted on Prosper.com between June 2006 and May 2007. The

summary statistics for each variable are reported separately for all loan listings and the set of loan listings that eventually funded. The "Credit

Grade", "Loan Information", and "Other Information Provided" categories provide information that was obtained directly from variables

provided by Prosper.com. The "Purpose of Loan" category and "Information from Pictures" category was coded by us using the descriptions and

pictures that individuals posted as part of their loan listings.

Table 1. Summary Statistics

For Pictures with Adults:

Gender

Main Content of Picture

Information from Listings Information from Pictures (for those with a picture)

OLS Logit OLS Logit

(1) (2) (3) (4)

Mean of Dependent Variable 0.093 0.093 Loan Purpose (BG: Unclear)Consolidate or Pay Debt 0.005 0.004

Single Female 0.004 0.011 (0.003)* (0.002)*(0.004) (0.003)*** Business/Entrepreneurship -0.015 -0.006

Couple -0.007 -0.001 (0.004)*** (0.003)**(0.004) (0.003) Pay Bills -0.015 -0.010

Group -0.011 -0.004 (0.007)** (0.006)(0.006)* (0.004) Education Expenses 0.001 0.001

Race (BG: White) (0.007) (0.005)Black -0.032 -0.024 Medical/Funeral Expenses -0.013 -0.014

(0.003)*** (0.003)*** (0.007)* (0.006)**Asian 0.002 0.004 Home Repairs 0.018 0.005

(0.009) (0.006) (0.010)* (0.006)Hispanic -0.018 -0.006 Auto Purchase -0.009 -0.005

(0.008)** (0.005) (0.008) (0.006)Age (BG: 35-60 yrs) Home/Land Purchase -0.023 -0.015

Less than 35 yrs 0.009 0.004 (0.008)*** (0.006)***(0.003)*** (0.002)* Auto Repairs -0.019 -0.015

More than 60 yrs -0.023 -0.011 (0.012) (0.007)**(0.011)** (0.007)* Luxury Item Purchase -0.013 -0.011

Happiness (BG: Neutral) (0.012) (0.008)Happy 0.007 0.002 Wedding -0.005 -0.006

(0.003)** (0.002) (0.012) (0.007)Unhappy -0.018 -0.016 Reinvest in Prosper 0.034 -0.010

(0.010)* (0.009)* (0.021) (0.006)*Weight (BG: Not Overweight) Taxes 0.010 0.008

Somewhat overweight 0.001 0.003 (0.019) (0.011)(0.004) (0.003) Vacation or Trip 0.032 0.006

Very overweight -0.016 -0.008 (0.020) -(0.011)(0.008)** (0.006) Multiplie of Above Reasons -0.004 -0.003

Attractiveness (BG: Average) (0.005) (0.003)Very attractive 0.007 0.004 Picture Characteristics X X

(0.008) (0.005) Month Fixed Effects X X

Very unattractive -0.002 -0.005 Credit Controls X X

(0.009) (0.007) R-Squared 0.31

Misc. Adult Information Observations 110,333 110,332Profesionally Dressed 0.000 0.002

(0.005) (0.003)Child With Adult in Picture -0.005 0.001

(0.003) (0.002)Signs of Military Involvement 0.014 0.025

(0.011) (0.009)***

* significant at 10%; ** significant at 5%; *** significant at 1%

Table 2. The Effect of Borrower Characteristics and Purpose on Loans Being Funded

Dependent Variable: Indicator = 1 if the Loan was Funded

Gender (BG: Single Male)

Notes: Coefficient values and standard errors clustered by borrower are presented using an OLS regression (Columns (1) and (3)) and a Logit regression (Columns

(2) and (4)) - marginal effects reported. The dependent variable in both regressions is a dummy variable indicating whether a particular loan listing was funded.

Each characteristic type for which a coefficient value is reported can be interpreted relative to its base group which is indicated in parenthesis. The coefficients

on other variables that are included in the regression (credit controls, month fixed effects, etc.) are omitted due to space constraints. The entire set of variables

used in these regressions is provided in the text under the heading "Baseline Regression Estimates".

(1) (2) (3) (4) (5) (6) (7) (8)

Mean of Dependent Variable 0.093 0.093 0.093 0.093 0.093 0.093 0.093 0.093

Race (BG: White)Black -0.051 -0.028 -0.027 -0.022 -0.021 -0.021 -0.024 -0.024

(0.002)*** (0.003)*** (0.002)*** (0.003)*** (0.003)*** (0.003)*** (0.003)*** (0.003)***Asian 0.010 0.000 0.006 0.006 0.007 0.006 0.005 0.004

(0.008) (0.006) (0.006) (0.006) (0.006) (0.006) (0.006) (0.006)Hispanic -0.028 -0.013 -0.012 -0.007 -0.006 -0.005 -0.006 -0.006

(0.006)*** (0.006)** (0.006)** (0.006) (0.006) (0.005) (0.005) (0.005)Credit Grade (BG: HR & NC)

AA 0.745(0.004)***

A 0.704(0.004)***

B 0.624(0.004)***

C 0.477(0.004)***

D 0.315(0.004)***

E 0.106(0.003)***

Other Key Credit VariablesMaximum Borrower's Rate 1.756

(0.022)***Debt to Income Ratio -0.014

(0.001)***$ Requested (thousands) -0.000

(.000)***

X X X X X X

X X X X X X

All other Credit Controls X X X X X

Long Description Text Controls X X X X

Month Fixed Effects X X X

Other Picture Characteristics X X

Loan Purpose Fixed Effects X

Observations 110,333 110,333 110,333 110,333 110,333 110,333 110,333 110,333

* significant at 10%; ** significant at 5%; *** significant at 1%

Cubic of Borrower's Max

Rate x Credit Grades

Cubic of Debt to Income

Ratio and Amount

Dependent Variable: Indicator = 1 if the Loan was Funded

Table 3. The Effect of Race on Loans Being Funded - Specification Robustness

Notes: Coefficient values and standard errors clustered by borrower are presented using Logit regressions - marginal effects reported. The dependent variable in all regressions is a dummy variable

indicating whether a particular loan listing was funded. Each column progressively includes a larger set of control variables. The coefficients on these control variables are omitted due to space

constraints. The entire set of variables used in these regressions is provided in the text under the heading "Baseline Regression Estimates".

AA & A B & C D & E HR & NC First 6 Months Last 6 Months

(1) (2) (3) (4) (5) (6)

Mean of Dependent Variable 0.335 0.225 0.108 0.035 0.087 0.096

Race (BG: White)

Black -0.041 -0.054 -0.033 -0.013 -0.020 -0.026(0.031) (0.014)*** (0.005)*** (0.002)*** (0.004)*** (0.003)***

Asian 0.011 0.000 -0.004 0.006 0.002 0.005(0.027) (0.019) (0.011) (0.006) (0.011) (0.006)

Hispanic 0.072 0.005 -0.011 -0.006 -0.022 0.006(0.054) (0.026) (0.010) (0.004) (0.008)*** (0.007)

Other Picture Characteristics X X X X X X

X X X X X X

Month Fixed Effects X X X X X X

Credit Controls X X X X X X

Observations 5,587 12,123 32,154 60,391 45,941 64,386

Panel B

Low Quartile 2nd Quartile 3rd Quartile High Quartile No College College

(1) (2) (3) (4) (5) (6)

Mean of Dependent Variable 0.059 0.072 0.091 0.147 0.086 0.116

Race (BG: White)

Black -0.021 -0.022 -0.022 -0.034 -0.019 -0.032

(.004)*** (0.004)*** (0.005)*** (0.007)*** (0.004)*** (0.006)***

Asian -0.003 -0.001 0.008 0.016 -0.006 0.003

(.009) (0.010) (0.011) (0.013) (0.008) (0.010)

Hispanic -0.020 0.000 0.003 -0.013 -0.003 -0.012

(.006)*** (0.009) (0.013) (0.013) (0.007) (0.014)Other Picture Characteristics X X X X X X

X X X X X X

Month Fixed Effects X X X X X X

Credit Controls X X X X X X

Observations 28,480 26,054 27,244 27,288 56,208 20,432

* significant at 10%; ** significant at 5%; *** significant at 1%

Notes: Coefficient values and standard errors clustered by borrower are presented using Logit regressions - marginal effects reported. The dependent variable in

all regressions is a dummy variable indicating whether a particular loan listing was funded. Columns (1) - (4) of Panel A present results from regressions using data

cut by credit grades. The second half of Panel A presents results from regressions using data from June 2006 to November 2006 (Column (5)) and December 2006

to May 2007 (Column (6)). Columns (1) - (4) of Panel B present results from regression using data cut by income quartiles. The second half of Panel A present

results from regressions using data from individuals whose self-reported occupation does not typically require a college degree (Column (5) and for those whose

occupation typically does require a college degree (Column (6)). The coefficients on other variables that are included in the regression (credit controls, month fixed

effects, etc.) are omitted due to space constraints. The entire set of variables used in these regressions is provided in the text under the heading "Baseline

Regression Estimates".

Loan Purpose Fixed Effects

Dependent Variable: Indicator = 1 if the Loan was Funded

Loan Purpose Fixed Effects

Table 4. The Effect of Race on Loans Being Funded - Sample Cuts

Sample Cut by Credit Grades Sample Cut by Time

Panel A

Sample Cut by Income Quartiles Sample Cut by Occupation

(1) (2) (3) (4)

Mean of Dependent Variable 0.182 0.161 Loan Purpose (BG: Unclear)Consolidate or Pay Debt -0.002 -0.002

Single Female -0.004 -0.004 (0.001)* (0.001)(0.001)*** (0.001)*** Business/Entrepreneurship 0.002 0.001

Couple -0.001 0.000 (0.001)* (0.001)(0.001) (0.001) Pay Bills 0.007 0.006

Group 0.001 0.001 (0.004)* (0.006)(0.002) (0.002) Education Expenses 0.003 0.002

Race (BG: White) (0.002) (0.003)Black 0.006 0.008 Medical/Funeral Expenses 0.005 0.002

(0.002)*** (0.002)*** (0.003)* (0.004)Asian 0.002 0.000 Home Repairs -0.001 -0.003

(0.002) (0.002) (0.002) (0.002)Hispanic 0.002 0.001 Auto Purchase 0.001 -0.001

(0.003) (0.003) (0.003) (0.003)Age (BG: 35-60 yrs) Home/Land Purchase 0.000 0.002

Less than 35 yrs -0.001 0.000 (0.003) (0.003)(0.001) (0.001) Auto Repairs 0.004 0.006

More than 60 yrs 0.000 0.003 (0.004) (0.005)(0.003) (0.004) Luxury Item Purchase -0.001 -0.001

Happiness (BG: Neutral) (0.003) (0.004)Happy -0.001 -0.001 Wedding 0.010 0.007

(0.001) (0.001) (0.004)** (0.005)Unhappy 0.002 0.002 Reinvest in Prosper 0.004 0.004

(0.004) (0.005) (0.002)** (0.002)**Weight (BG: Not Overweight) Taxes -0.007 -0.012

Somewhat overweight 0.002 0.002 (0.004)* (0.005)**(0.001)* (0.001) Vacation or Trip 0.005 0.007

Very overweight 0.003 0.005 (0.004) (0.008)(0.002) (0.003) Multiplie of Above Reasons 0.003 0.003

Attractiveness (BG: Average) (0.002) (0.002)Very attractive 0.003 0.003 Open Funding Option Only X

(0.002)* (0.002) Picture Characteristics X X

Very unattractive 0.006 0.009 Month Fixed Effects X X

(0.003)** (0.003)*** Credit Controls X X

Misc. Adult Information R-Squared 0.79 0.76Profesionally Dressed 0.003 0.003 Observations 10,207 6,419

(0.001)** (0.001)*Child With Adult in Picture 0.003 0.002

(0.001)*** (0.001)Signs of Military Involvement -0.002 0.000

(0.003) (0.004)

* significant at 10%; ** significant at 5%; *** significant at 1%

Notes: Coefficient values and standard errors clustered by borrower are presented using two OLS regressions. The dependent variable in both regressions is the

final interest rate that borrowers have to pay for a particular loan. The regression presented in Columns (1) and (3) uses the entire sample of funded loans while

the regression reported in Columns (2) and (4) restricts the sample to loans for which the setting of the loan listing was such to allow an auction system to

determine the final interest rate. Each characteristic type for which a coefficient value is reported can be interpreted relative to its base group which is indicated

in parenthesis. The coefficients on other variables that are included in the regression (credit controls, month fixed effects, etc.) are omitted due to space

constraints. The entire set of variables used in these regressions is provided in the text under the heading "Baseline Regression Estimates".

Table 5. The Effect of Borrower Characteristics and Purpose on the Final Interest Rate for Funded Loans

Dependent Variable: The Final Interest Rate for Funded Loans

OLS OLS

Gender (BG: Single Male)

Current or

Paid Off

1 Month

Late

2 Months

Late

3 Months

Late

4+ Months

Late

Total # of

Loans

0.78 0.02 0.02 0.02 0.17 10,118

7 0.91 0.02 0.04 0.01 0.02 230

8 0.86 0.03 0.02 0.02 0.07 1,073

9 0.84 0.03 0.02 0.01 0.10 1,095

10 0.84 0.02 0.01 0.02 0.11 1,153

11 0.77 0.02 0.02 0.03 0.16 912

12 0.76 0.02 0.02 0.02 0.19 982

13 0.74 0.02 0.02 0.02 0.20 950

14 0.73 0.01 0.02 0.02 0.22 701

15 0.72 0.01 0.02 0.02 0.24 704

16 0.76 0.02 0.01 0.01 0.20 613

17 0.71 0.01 0.01 0.02 0.24 753

18 0.74 0.01 0.01 0.01 0.24 560

19 0.71 0.02 0.01 0.01 0.25 392

Race

White 0.79 0.02 0.02 0.02 0.15 3,756

Black 0.63 0.03 0.02 0.03 0.29 533

Asian 0.80 0.02 0.02 0.02 0.14 163

Hispanic 0.69 0.02 0.04 0.04 0.21 103

Notes: Summary statistics are provided for loan performance broken down by loan maturity and race for all 10,118 loan listings

between June 2006 and May 2007 that were fully funded. This loan performance data was provided by Prosper.com in December

2007. For each loan type, the fraction of loans that were current or paid off, 1 month late, 2 months late, 3 months late, and 4+

months late is reported. The total # of loans for each loan type is also reported.

Table 6. Loan Performance Summary Statistics

Current Status of All Funded Loans (Fractions Reported)

Age of the Loan

(Months)

All Loans

Default

Return Type

I

Return Type

II

Return Type

III

Return Type

I

Return Type

II

Return Type

III

1.047 1.066 1.084 1.081 1.099 1.116

Single Female 0.139 -0.023 -0.016 -0.016 -0.032 -0.023 -0.020(0.094) (0.016) (0.016) (0.016) (0.018)* (0.018) (0.017)

Couple -0.064 0.001 0.002 0.003 0.001 -0.002 -0.001(0.110) (0.016) (0.016) (0.016) (0.017) (0.017) (0.016)

Group 0.077 0.001 -0.001 0.005 0.023 0.026 0.049(0.161) (0.026) (0.026) (0.025) (0.026) (0.025) (0.023)**

Race (BG: White)

Black 0.346 -0.086 -0.082 -0.084 -0.075 -0.073 -0.073(0.100)*** (0.023)*** (0.023)*** (0.023)*** (0.027)*** (0.027)*** (0.026)***

Asian -0.230 0.017 0.017 0.015 0.017 0.020 0.020(0.210) (0.034) (0.033) (0.032) (0.034) (0.032) (0.030)

Hispanic 0.050 -0.050 -0.051 -0.031 -0.060 -0.065 -0.050(0.231) (0.047) (0.047) (0.046) (0.057) (0.057) (0.054)

Age (BG: 35-60 yrs)

Less than 35 yrs -0.008 0.000 -0.003 -0.003 0.020 0.021 0.019(0.082) (0.014) (0.013) (0.013) (0.015) (0.014) (0.014)

More than 60 yrs -0.042 -0.037 -0.039 -0.012 -0.060 -0.058 -0.010(0.362) (0.044) (0.043) (0.041) (0.054) (0.053) (0.048)

Happiness (BG: Neutral)

Happy -0.072 0.022 0.017 0.012 0.012 0.005 0.001(0.083) (0.015) (0.014) (0.014) (0.016) (0.015) (0.015)

Unhappy 0.443 -0.016 -0.039 -0.070 0.036 0.017 -0.019(0.266)* (0.060) (0.060) (0.060) (0.070) (0.071) (0.071)

Weight (BG: Not Overweight)

Somewhat overweight 0.116 -0.021 -0.017 -0.008 -0.014 -0.006 0.011(0.095) -0.017 (0.017) (0.017) (0.020) (0.019) (0.018)

Very overweight -0.164 0.072 0.052 0.052 0.029 0.010 0.005(0.207) (0.039)* (0.039) (0.040) (0.045) (0.045) (0.047)

Attractiveness (BG: Average)

Very attractive -0.097 -0.012 -0.014 0.012 -0.026 -0.022 0.004(0.181) (0.030) (0.030) (0.028) (0.034) (0.032) (0.030)

Very unattractive 0.309 -0.049 -0.046 -0.056 -0.011 0.003 -0.005(0.234) (0.048) (0.046) (0.045) (0.055) (0.051) (0.049)

Misc. Adult Information

Profesionally Dressed 0.129 -0.025 -0.013 -0.012 0.007 0.013 0.013(0.117) (0.018) (0.018) (0.017) (0.020) (0.019) (0.018)

Child With Adult in Picture 0.130 -0.035 -0.029 -0.022 -0.038 -0.033 -0.029(0.080) (0.014)** (0.014)** (0.013)* (0.015)** (0.015)** (0.014)**

Signs of Military Involvement 0.458 -0.034 -0.029 -0.052 0.019 0.025 -0.005(0.260)* (0.048) (0.047) (0.046) (0.051) (0.049) (0.048)

X X X X X X X

Month Fixed Effects X X X X X X X

Credit Controls X X X X X X X

R-Squared 0.26 0.26 0.26 0.27 0.28 0.27

Observations 9,963 10,113 10,113 10,113 6,369 6,369 6,369

* significant at 10%; ** significant at 5%; *** significant at 1%

Table 7. The Effect of Borrower Characteristics on Net Return on Investment

Notes: Coefficient values and standard errors clustered by borrower are presented using a Cox proportional hazard model (Column (1)) and OLS regression (Columns (2) - (7)). The

dependent variable for Column (1) is a default indicator and hazard ratios are reported as coefficients. The dependent variable for Columns (2) - (7) is the 3-year net return on a $1

investment into a particular loan. This 3-year net return was calculated using three separate assumptions of default. Definitions for each assumption can be found in the text. Columns (2)

- (4) use the entire sample of funded loans while Columns (5) - (7) restrict the sample to loans for which the setting of the loan listing was such to allow an auction system to determine the

final interest rate. Each characteristic type for which a coefficient value is reported can be interpreted relative to its base group which is indicated in parenthesis. The coefficients on other

variables that are included in the regression (credit controls, month fixed effects, etc.) are omitted due to space constraints. The entire set of variables used in these regressions is

provided in the text under the heading "Baseline Regression Estimates".

All Funded Loans Only Open-Auction Loans

Gender (BG: Single Male)

Loan Purpose Fixed Effects

Mean of Dependent Variable

OLS -- Dependent Variable: 3-Year Return on Each Dollar Invested by Return TypeHazard Model - Dep

Var: Default

Figure A1 b Standardized Bias for Key Variables

Figure A1.a. Overlap in the Distribution of Estimated Propensity Score of being Black

Notes: Figure A1.a. displays the histogram of the estimated propensity score of being black for blacks and whites. The propensity score is thepredicted probability that a listing has a black picture from a flexibly specified logit estimation. See the text for further description. A smallnumber of listings (99 listings) with black pictures had estimated propensity scores that were off the support (i.e., above the maximumpropensity score for whites); these listings are shown in the green bar at the right tail of the black histogram. Figure A1.b. displays thestandardized bias (see Rosenbaum and Rubin (1985)) for a number of key variables both before and after matching. For each covariate thismeasures is calculated as the difference of sample means between blacks and whites as a percentage of the square root of the average ofsample variances in both groups. The matching is done using nearest neighbors with a caliper of 0.01. The reduction in standardized biasshows the effect of the matching procedure.

couple

revolving credit balance

credit grade C

credit grade AA

$ amount requested

credit grade D

credit grade B

child also in picture

bank‐card utilization

length of description

credit grade A

group member

credit grade E

debt to income ratio

text readibility score

credit grade NC

income

borrower's max int. rate

no long description 

% words mispelled

professionally dressed

delinquencies last 7 years

current delinquencies

credit grade HR

female   

‐60 ‐40 ‐20 0 20 40 60

Figure A1.b. Standardized Bias for Key Variables      (Black ‐White) Before and After Matching 

unmatched

matched